The company has announced the results of using its Claude Mythos Preview language model to find vulnerabilities in cryptographic algorithms. The study identified two cryptanalytic attacks aimed at the HAWK post-quantum digital signature scheme and an abbreviated seven-round version of the AES-128 standard. Representatives of the organization stressed that these achievements are purely theoretical in nature and have no practical impact on modern secure systems. Nevertheless, the results demonstrate the ability of advanced artificial intelligence models to conduct sophisticated cryptographic research at a level comparable to the work of qualified cryptographic experts.

In the case of the HAWK algorithm, which is considered as a candidate for the standardization of post-quantum digital signatures, the model revealed a previously unknown mathematical symmetry. This discovery made it possible to reduce the assessment of the effective durability of keys, which theoretically may require an increase in their size to maintain the previous level of protection. As for the AES-128 standard, the researchers described the Möbius Bridge method, which accelerates the attack on the shortened version of the cipher. However, this scenario remains impractical, as it assumes a plaintext selection model with a huge number of queries, which makes it inapplicable against full versions of the algorithm in real-world conditions.

To systematically evaluate such opportunities, the company, together with academic partners, presented a specialized CryptanalysisBench benchmark. This tool includes tasks for various families of cryptographic primitives, requiring models not just to provide a theoretical explanation, but to provide a working attack script. Testing has shown that advanced models successfully cope with entry-level tasks, but their effectiveness drops sharply when working with full-fledged schemes without published vulnerabilities. These studies highlight the growing role of artificial intelligence in security auditing, while pointing to the need for rigorous expert verification of generated algorithmic solutions.